Csci 2111: Data and File Structures Week 8, Lectures 1 & 2

Csci 2111: Data and File StructuresWeek 8, Lectures 1 & 2 Multi-Level Indexing and B-Trees

Statement of the Problem • When indexes grow too large they have to be stored on secondary storage. • However, there are two fundamental problems associated with keeping an index on secondary storage: • Searching the index must be faster than binary searching. • Insertion and deletion must be as fast as search.

Indexing with Binary Search Trees: Negative Aspects • A sorted list can be expressed in a Binary Search Tree representation. • However, there are 2 problems with binary search trees: • They are not fast enough for disk resident indexing. • There is no effective strategy of balancing the tree. We will look at 2 solutions: AVL Trees and Paged Binary Trees.

Indexing with Binary Search Trees: Positive Aspects • Tree structures give us an important new capability: we no longer have to sort the file to perform a binary search. • To add a new key, we simply link it to the appropriate leaf node. • If the tree remains balanced, then the search performance on this tree is good. • Problems occur when the tree gets unbalanced.  We will look for schemes that allow trees to remain balanced

AVL Trees I • AVL Trees allow us to re-organize the nodes of the tree as we receive new keys, maintening a near optimal tree structures. • An AVL Tree is a height-balanced tree, i.e., a tree that places a limit on the amount of difference allowed between the heights of any two sub-trees sharing a common root. • In an AVL or HB-1 tree, the maximum allowable difference is one.

AVL Trees II • The two features that make AVL trees important are: • By setting a maximum allowable difference in the height of any two sub-trees, AVL trees guarantee a minimum level of performance in searching. • Maintaining a tree in AVL form as new nodes are inserted involves the use of one of a set of four possible rotations. Each of the rotations is confined to a single local area of the tree. The most complex of the rotations requires only five pointer reassignments.

AVL Tree III • AVL Trees are not, themselves, directly applicable to most file structures because like all strictly binary trees, they have too many levels--they are too deep. • AVL Trees, however, are important because they suggest that it is possible to define procedures that maintain height-balance. • AVL Trees’ search performance approximates that of a completely balanced tree. For a completely balanced tree, the worst-case search to find a key is log2(N+1). For an AVL Tree it is 1.44 Log2(N+2).

Paged Binary Trees • AVL trees tackle the problem of keeping an index in sorted order cheaply. They do not address the problem regarding the fact that Binary Searching requires too many seeks. • Paged Binary trees addresses this problem by locating multiple binary nodes on the same disk page. • In a paged system, you do not incur the cost of a disk seek just to get a few bytes. Instead, once you have taken the time to seek to an area of the disk, you read in an entire page from the file. • When searching a Binary Tree, the number of seeks necessary is log2(N+1). It is logk+1(N+1) in the paged version.

Problems with Paged Trees I • Inefficient Disk Usage • How should we build a paged tree? • Easy if we know what the keys are and their order before starting to build the tree. • Much more difficult if we receive keys in random order and insert them as soon as we receive them. The problem is that the wrong keys may be placed at the root of the trees and cause an imbalance.

Problems with Paged Trees II • Three problems arise with paged trees: • How do we ensure that the keys in the root page turn out to be good separator keys, dividing up the set of other keys more or less evenly. • How do we avoid grouping keys that shouldn’t share a page? • How can be guarantee that each of the pages contains at least some minimum number of keys?

Multi-Level Indexing: A Better Approach to Tree Indexes • Up to this point, in this lecture, we’ve looked at indexing a file based on building a search tree. However, there are problems with this approach (see previous slide). • Instead, we get back to the notion of the simple indexes we saw earlier in the course, but we extend this notion to that of multi-record indexes and then, multi-level indexes. • While multi-record multi-level indexes really help reduce the number of disk accesses and their overhead space costs are minimal, inserting a new key or deleting an old one is very costly.

B-Trees: Addressing the problems of Paged Trees and Multi-Level Indexing • Trees appear to be a good general solution to indexing, but each particular solution we’ve looked at so far presents some problems. • Paged Trees suffer from the fact that they are built downward from the top and that a “bad” root may unbalance the construct. • Multi-Level Indexing takes a different approach that solves many problems but creates costly insertion and deletion. • An ideal solution would be one that combines the advantages of the previous solutions and does not suffer from their disadvantages. • B-Trees appear to do just that!

B-Trees: An Overview • B-Trees are built upward from the bottom rather than downward from the top, thus addressing the problems of Paged Trees: with B-Trees, we allow the root to emerge rather than set it up and then find ways to change it. • B-Trees are multi-level indexes that solve the problem of linear cost of insertion and deletion. • B-Trees are now the standard way to represent indexes.

P W D M P T W A N P B C D U W I M R S T G Example of a B-Tree Note: references to actual record only occur in the leaf nodes. The interior nodes are only higher level indexes (this is why there are duplications in the tree)

How do B-Trees work? Main Ideas • Each node of a B-Tree is an Index Record. Each of these records has the same maximum number of key-reference pairs called the order of the B-Tree. The records also have a minimum number of key-reference pairs, typically, half the order. • When inserting a new key into an index record that is not full, we simply need to update that record and possibly go up the tree recursively. • When inserting a new key into an index record that is full, this record is split into two, each with half of the keys. The largest key of the split record is promoted which may cause a new recursive split.

W P M D P T W U A B C D N P W G I M R S T Searching a B-Tree • Problem 1: Look for L • Problem 2: Look for S

Insertion into a B-Tree: General Strategy • Search all the way down to the leaf level in order to find the insertion location. • Insertion, overflow detection, and splitting on the upward path. • Creation of a new root node if the current root was split.

S T C D D T A C D S T Insertion into a B-Tree: No Split & Contained Splits After inserting C, S, D, T: Inserting A

D M P W A B C D G I M N P S T U W W P M D P W T A B C D N P U W G I M R S T Insertion into a B-Tree: Recursive Split Inserting R

Formal Definition of B-Tree Properties In a B-Tree of order m, • Every page has a maximum of m descendants • Every page, except for the root and leaves, has at least m/2 descendants. • The root has at least two descendants (unless it is a leaf). • All the leaves appear on the same level. • The leaf level forms a complete, ordered index of the associated data file.

Worst-Case Search Depth I • Given 1,000,000 keys and a B-Tree of order 512, what is the maximum number of disk accesses necessary to locate a key in the tree? In other words, how deep will the tree be? • Each key appears in the leaf ==> What is the maximum height of a tree with 1,000,000 leaves? • The maximum height will be reached if all pages (or nodes) in the tree has the minimum allowed number of descendents • For a B-Tree of order m, the minimum number of descendents from the root page is 2. It is m/2 for all the . other pages.

Worst-Case Search Depth II • For any level d of a B-Tree, the minimum number of descendants extending from that level is . 2 m/2 d-1 • For a tree with N keys in its leaves, we have N 2 m/2 d-1  d  1 + logm/2 (N/2) • For m= 512 and N= 1,000,000, we thus get d  3.37

Deletion from a B-Tree: Rules for Deleting a key k from a node n • If n has more than the number of keys and the k is not the largest in n, simply delete k from n. • If n has more than the minimum number of keys and the k is the largest in n, delete k and modify the higher level indexes to reflect the new largest key in n. • If n has exactly the minimum number of keys and one of the siblings of n has few enough keys, merge n with its sibling and delete a key from the parent node. • If n has exactly the minimum number of keys and one of the siblings of n has extra keys, redistribute by moving some keys from a sibling to n, and modify the higher level indexes to reflect the new largest keys in the affected nodes.

I P Z D G I M P T X Z J K L M A B C D Q R S T Y Z E F G H I N O P U V W X Deletion from a B-Tree: Example • Problem 1: Delete C • Problem 2: Delete P • Problem 3: Delete H

Redistribution during Insertion • Redistribution during insertion is a way to avoid, or at least postpone, the creation of new pages. • Redistribution allows us to place some of the overflowing keys into another page instead of splitting an overflowing page. • B* Trees formalize this idea

Properties of a B* Tree • Every page has a maximum of m descendants. • Every page except for the root has at least  (2m-1)/3 descendants. • The root has at least two descendants (unless it is a leaf) • All the leaves appear on the same level. The main difference between a B-Tree and a B* Tree is in the second rule.

Csci 2111: Data and File Structures Week 8, Lectures 1 & 2